Method for training an image recognition deep learning model

Abstract

A training method of an image recognition deep learning model includes selecting and labeling part of original data from a plurality of original data; training the labeled data according to a deep learning algorithm to generate an image recognition deep learning model; inputting the part of original data to the image recognition deep learning model to infer an inference result and calculate a correct rate, and then determining whether the correct rate is greater than or equal to a threshold value. When the correct rate is greater than or equal to the threshold value, a final image recognition deep learning model is obtained; when the correct rate is less than the threshold value, the original data with the incorrect inference result is relabeled to generate relabeled data and the image recognition deep learning model is updated, and then the foregoing process is repeated until the correct rate is greater than or equal to the threshold value. In this way, the image recognition deep learning model can be quickly converged.

Description

Technical Field

[0001] This disclosure relates to a training method for a deep learning model, and more particularly to a training method for a supervised learning model that can be applied to image recognition and converges quickly. Background Technology

[0002] In deep learning and machine learning, supervised learning is widely recognized and supported by numerous experimental data, as it currently boasts the highest accuracy rate. However, its application in image recognition requires a vast amount of labeled and classified images to achieve a high accuracy rate for the recognition model. Therefore, this aforementioned drawback necessitates additional human resources to label and classify a large number of raw images before training the recognition model, significantly increasing the development time required for training the model.

[0003] In view of this, the development of a training method for supervised learning models that can be applied to image recognition and converge quickly is something that the public eagerly anticipates, and it is also a goal and direction that relevant industry players must strive to achieve through research and development. Summary of the Invention

[0004] The purpose of this disclosure is to provide a training method for an image recognition deep learning model. This method selects a small amount of raw data from a large amount of raw data and models an image recognition deep learning model. Then, it inputs untrained raw data into the image recognition deep learning model and re-labels the raw data with incorrect inference results to repeatedly train and update the image recognition deep learning model. This allows the final image recognition deep learning model to converge quickly while maintaining a high accuracy rate.

[0005] According to one embodiment of this disclosure, a method for training an image recognition deep learning model is provided. This method obtains a final image recognition deep learning model and includes a first data selection step, a data labeling step, a model training step, a second data selection step, and a data inference step. The first data selection step includes selecting N data points from T original data points. i Given _n_ original data, where i = 1 to n, and T, N i All of i and n are positive integers. The data marking step involves setting N... i The original data are labeled to generate N. i Labeled data. The model training steps include training N based on a deep learning algorithm. i The first step involves using labeled data to generate an i-th image recognition deep learning model. The second data selection step includes selecting data from TM... i N are selected from the original data.i+1 There are M original data points. i The input is a positive integer. The data inference steps include input P. i The original data is fed into the i-th image recognition deep learning model to infer P. i One inference result, and based on P i The inference results are used to calculate the accuracy of the i-th image recognition deep learning model, and then it is determined whether the accuracy is greater than or equal to a threshold value, where P i Let be a positive integer. When the accuracy is greater than or equal to the threshold, the i-th image recognition deep learning model is the final image recognition deep learning model. When the accuracy is less than the threshold, a data relabeling step and a model update step are executed, then i is set to i+1, and the second data selection step and data inference step are repeated until the accuracy is greater than or equal to the threshold. The data relabeling step includes relabeling P... i At least one of the original data is used to generate at least one relabeled data. The model update step includes retraining the at least one relabeled data based on a deep learning algorithm to update the i-th image recognition deep learning model. i+1 The original data is different from N. i The original data. The data inference steps further include obtaining the corresponding P from the storage unit. i P of the original data i A real result, and then based on P i Calculate P based on the actual results. i The accuracy is derived from the inference results. In the data relabeling step, P... i The inference result corresponding to at least one of the original data does not match the actual result. Attached Figure Description

[0006] Figure 1 This is a schematic diagram illustrating a training system for an image recognition deep learning model according to an embodiment of the present disclosure;

[0007] Figure 2 This is a flowchart illustrating a training method for an image recognition deep learning model according to a first embodiment of the present disclosure; and

[0008] Figure 3 This is a flowchart illustrating a training method for an image recognition deep learning model according to a second embodiment of the present disclosure.

[0009] The reference numerals in the attached figures are explained as follows:

[0010] 100: Training System for Deep Learning Models for Image Recognition

[0011] 110: Storage unit

[0012] 112: Raw Data

[0013] 114: Deep Learning Algorithms

[0014] 116: Threshold value

[0015] 118: Actual Result

[0016] 120: Processing Unit

[0017] 121: Tag Data

[0018] 122: The first deep learning model for image recognition

[0019] 1221, 1251: Accuracy

[0020] 123, 126: Inference Results

[0021] 124: Relabel data

[0022] 125: Second Image Recognition Deep Learning Model

[0023] 200, 300: Training Methods for Deep Learning Models for Image Recognition

[0024] S210, S310: First data selection step

[0025] S220, S320: Data tagging steps

[0026] S230, S330: Model Training Steps

[0027] S240, S340: Second data selection steps

[0028] S250, S350: Data Inference Steps

[0029] S260, S360: Model Determination Steps

[0030] S262, S362: Data Relabeling Steps

[0031] S264, S364: Model Update Steps Detailed Implementation

[0032] Please refer to Figure 1 This is a schematic diagram illustrating a training system 100 for an image recognition deep learning model according to an embodiment of the present disclosure. Figure 1 It is understood that the training system 100 for the image recognition deep learning model includes a storage unit 110 and a processing unit 120, and can be various intelligent devices and / or computer devices for running the image recognition deep learning model.

[0033] Storage unit 110 stores T units of raw data 112, a deep learning algorithm 114, and a threshold value 116, where T is a positive integer. The deep learning algorithm 114 can be a supervised learning algorithm, but this disclosure is not limited to this. Furthermore, the threshold value 116 is used to assess the accuracy of the image recognition deep learning model trained based on the deep learning algorithm 114, and the threshold value 116 can be configured according to the user's needs. Specifically, storage unit 110 is a machine-readable medium, which can be, but is not limited to, random access memory, read-only memory, flash memory, hard disk, or a combination of the aforementioned devices, and can be used to store the aforementioned image recognition deep learning model.

[0034] The processing unit 120 is coupled to the storage unit 110, and may be, but is not limited to, a general-purpose processor, a microcontroller, or the like. In the embodiments disclosed herein, the processing unit 120 may access modules and program code / electronic instructions recorded in the storage unit 110 to implement the training method of the image recognition deep learning model proposed in this disclosure, the details of which are described below.

[0035] Please refer to Figure 2 This is a flowchart illustrating a training method 200 for an image recognition deep learning model according to a first embodiment of the present disclosure. The training method 200 for the image recognition deep learning model is used to obtain a final image recognition deep learning model and includes a first data selection step S210, a data labeling step S220, a model training step S230, a second data selection step S240, and a data inference step S250, and can be performed by… Figure 1 The training system 100 for the image recognition deep learning model is executed, and the following is the configuration. Figure 1 Component description shown Figure 2 Details of each step.

[0036] The first data selection step S210 is to select N from T original data. i "T original data", which includes T original data 112 read by the drive processing unit 120 from the storage unit 110, and N original data 112 selected from the T original data 112. i There are 112 original data points, where i = 1 to n, T, N i i and n are both positive integers, and T is greater than N. i Furthermore, the first data selection step S210 may further include driving the arithmetic processing unit 120 to group the T raw data 112 into multiple raw data groups, and then selecting a portion of the raw data 112 from each raw data group to integrate N. iThere are 30,000 original data points 112. For example, T = 30,000. The processing unit 120 divides the 30,000 original data points 112 into 5 groups of original data points; in other words, each group of original data points contains the same number (i.e., 6,000) of original data points 112. Specifically, the processing unit 120 can group the 30,000 original data points 112 into 5 groups of original data points based on multiple classification features, so that each group of original data points has the same classification features. For example, the original data point 112 can be an original image of an animal (such as a cat or dog). In the same group of original data points, the animals in multiple original images can have the same size, breed, or color. In this way, by dividing the data into groups of equal size, the N in the subsequent data labeling step S220 can be made more accurate. i Each labeled data point will have the same number of data points for different features, which can accelerate the convergence speed of the final image recognition deep learning model.

[0037] Data labeling step S220 is "for N i "Labeling the original data", which includes driving the processing unit 120 to pair N i The original data 112 are labeled to generate N. i 1 tag data.

[0038] Model training step S230 is "generating the i-th image recognition deep learning model", which includes driving the computation processing unit 120 to train N according to the deep learning algorithm 114. i I labeled data to generate the i-th image recognition deep learning model.

[0039] The second data selection step S240 is "from TM" i N are selected from the original data. i+1 "raw data", which includes the driving processing unit 120 from TM i N is selected from the original data 112. i+1 There are 112 original data points, of which M i M is a positive integer and i Greater than N i+1 and satisfy the following conditions: It should be noted that N i+1 The original data 112 is different from N. i The original data is 112, that is, N. i+1 The original data 112 is the original data 112 that has not been trained by the deep learning algorithm 114.

[0040] The data inference step S250 is to "infer P". i "An inference result is obtained and the accuracy rate is determined to be greater than or equal to the threshold value," which includes input P to the driving arithmetic processing unit 120. iThe original data 112 are used to deduce P from the image recognition deep learning model. i One inference result, and based on P i The inference results are used to calculate the accuracy of the i-th image recognition deep learning model, and then it is determined whether the accuracy is greater than or equal to the threshold value of 116, where P i It is a positive integer and satisfies the following conditions: Furthermore, storage unit 110 can store T actual results (Groundtruths) 118, each corresponding to one of T original data. The data inference step S250 can further include driving the arithmetic processing unit 120 to obtain the corresponding P from the T actual results 118 in storage unit 110. i P of the original data i The actual result is 118, and then based on P i Calculate P based on 118 actual results. i The accuracy of the i-th image recognition deep learning model is derived from the inference result.

[0041] When the accuracy rate is greater than or equal to the threshold value of 116, the processing unit 120 continues to execute the model determination step S260. The model determination step S260 is "obtaining the final image recognition deep learning model," which includes driving the processing unit 120 to determine that the i-th image recognition deep learning model is the final image recognition deep learning model. When the accuracy rate is less than the threshold value of 116, the processing unit 120 continues to execute the data re-labeling step S262 and the model update step S264, then sets i to i+1 and repeats the second data selection step S240 and the data inference step S250 until the accuracy rate is greater than or equal to the threshold value of 116.

[0042] The data relabeling step S262 is "Relabel P" i "at least one of the original data to generate at least one relabeled data", which includes driving the arithmetic processing unit 120 to relabel P i At least one of the original data 112 is used to generate at least one relabeled data. It should be noted that in the data relabeling step S262, P... i The inference result corresponding to at least one of the original data 112 does not conform to the actual result 118; in other words, when input P i After the original data 112 is processed by the i-th image recognition deep learning model, the computation processing unit 120 will re-label the original data 112 with incorrect inference results.

[0043] The model update step S264 is "retraining at least one relabeled data to update the i-th image recognition deep learning model", which includes driving the operation processing unit 120 to retrain at least one relabeled data according to the deep learning algorithm 114 to update the i-th image recognition deep learning model.

[0044] Therefore, the training method 200 for the image recognition deep learning model disclosed herein selects a small number (N) of the data from a large number (T) of original data 112. i 112 original data points were used to model an image recognition deep learning model, and then the trained original data 112 (i.e., N) were used to generate the model. i The original data 112 and the untrained original data 112 (i.e., N) i+1 After the original data 112 is input into the image recognition deep learning model, the original data 112 with incorrect inference results are relabeled to repeatedly train and update the image recognition deep learning model, so that the final image recognition deep learning model can converge quickly and at the same time maintain a high accuracy.

[0045] Please refer to Figure 3 This is a flowchart illustrating a training method 300 for an image recognition deep learning model according to a second embodiment of the present disclosure. The training method 300 for the image recognition deep learning model is equivalent to... Figure 2 The training method 200 for the image recognition deep learning model operates in a scenario where n=2, and includes a first data selection step S310, a data labeling step S320, a model training step S330, a second data selection step S340, and a data inference step S350, and can be derived from... Figure 1 The training system 100 for the image recognition deep learning model is executed, and the following is the configuration. Figure 1 Component description shown Figure 3 Details of each step.

[0046] The first data selection step S310 includes the driving operation processing unit 120 dividing 30,000 (T = 30,000) raw data 112 into 5 raw data groups, and then selecting 60 raw data 112 from each raw data group to integrate 300 (N1 = 300) raw data 112. The data labeling step S320 includes the driving operation processing unit 120 labeling the 300 raw data 112 to generate 300 labeled data 121. The model training step S330 includes the driving operation processing unit 120 training the 300 labeled data 121 according to the deep learning algorithm 114 to generate the first image recognition deep learning model 122. The second data selection step S340 includes the driving operation processing unit 120 selecting 200 (N2 = 200) raw data 112 from 29,700 (T - M1 = T - N1 = 30,000 - 300) raw data 112. The data inference step S350 includes the driving operation processing unit 120 inputting 500 (N1+N2=300+200=500) raw data 112 to the first image recognition deep learning model 122 to infer 500 inference results 123, and calculating the 500 inference results 123 based on the 500 actual results 118 to generate the accuracy 1221 of the first image recognition deep learning model 122, and then determining whether the accuracy 1221 is greater than or equal to the threshold value 116.

[0047] In the second embodiment, the threshold value 116 can be set to 0.9, and the accuracy rate 1221 can be 0.88. Therefore, the accuracy rate 1221 is less than the threshold value 116 (i.e., no). The processing unit 120 then executes the data relabeling step S362 and the model update step S364. The data relabeling step S362 includes driving the processing unit 120 to relabel at least one piece of original data 112 that does not conform to the actual result 118 of the inference result 123 to generate at least one relabeled data 124. The model update step S364 includes driving the processing unit 120 to retrain at least one piece of relabeled data 124 according to the deep learning algorithm 114 to update the first image recognition deep learning model 122 into a second image recognition deep learning model 125. Next, the processing unit 120 executes the second data selection step S340 and the data inference step S350.

[0048] In the second data selection step S340, the arithmetic processing unit 120 selects 120 (N3 = 120) original data from 29,500 (T-M2 = T-(N1+N2) = 30,000-500) original data 112. In the second data inference step S350, the processing unit 120 inputs 620 (N1+N2+N3=300+200+120=620) raw data 112 to the second image recognition deep learning model 125 to infer 620 inference results 126. Based on the 620 actual results 118, the 620 inference results 126 are calculated to generate the accuracy 1251 of the second image recognition deep learning model 125. Then, it is determined whether the accuracy 1251 is greater than or equal to a threshold value 116, where the accuracy 1251 can be 0.98 and is greater than or equal to the threshold value 116 (i.e., yes). Therefore, the processing unit 120 continues to execute the model determination step S360. The model determination step S360 includes driving the processing unit 120 to determine that the second image recognition deep learning model 125 is the final image recognition deep learning model. Therefore, the training method 300 of the image recognition deep learning model disclosed herein updates the first image recognition deep learning model 122 by retraining and relabeling data 124, so that the final image recognition deep learning model can converge quickly and maintain a high accuracy, wherein the accuracy of the final image recognition deep learning model can be as high as 98%, but the total time spent can save 97.6% of the time compared with the existing supervised learning.

[0049] As can be seen from the above implementation methods, the present disclosure has the following advantages: First, labeling and classifying only a small amount of original data can greatly reduce the development time of the training model. Second, by averaging the original data, the labeled data (i.e., training data) will have the same amount for different features, thereby accelerating the convergence speed of the final image recognition deep learning model. Third, using relabeled original data with erroneous inference results to repeatedly train and update the image recognition deep learning model not only enables the final image recognition deep learning model to converge quickly, but also maintains a high accuracy rate.

[0050] Although the present disclosure has been described above with reference to embodiments, it is not intended to limit the present disclosure. Any person skilled in the art may make various modifications and alterations without departing from the spirit and scope of the present disclosure. Therefore, the scope of protection of the present disclosure shall be determined by the appended claims.

Claims

1. A training method for an image recognition deep learning model, used to obtain a final image recognition deep learning model, characterized in that, The training method for this image recognition deep learning model includes the following steps: The first data selection step includes selecting N data points from T original data points. i Given _n_ original data, where i = 1 to n, and T, N i i and n are all positive integers; A data labeling step, including the labeling of N i The original data are labeled to generate N. i One labeled data; A model training step includes training the N model using a deep learning algorithm. i I labeled data to generate an i-th image recognition deep learning model; The second data selection step includes selecting data from TM. i N are selected from the original data. i+1 There are M original data points. i are positive integers; and A data inference step, including input P i The original data is fed into the i-th image recognition deep learning model to infer P. i One inference result, and based on this P i The inference results are used to calculate the accuracy of the i-th image recognition deep learning model, and then it is determined whether the accuracy is greater than or equal to a threshold value, where P i It is a positive integer; Wherein, when the accuracy is greater than or equal to the threshold value, the i-th image recognition deep learning model is the final image recognition deep learning model; When the accuracy is less than the threshold value, a data relabeling step and a model update step are executed. Then, i is set to i+1 and the second data selection step and the data inference step are executed repeatedly until the accuracy is greater than or equal to the threshold value. The data relabeling step includes relabeling the P i At least one of the original data is used to generate at least one relabeled data; The model update step includes retraining the at least one labeled data according to the deep learning algorithm to update the i-th image recognition deep learning model; Among them, the N i+1 The original data is different from N. i One set of raw data; The data inference process further includes: Obtain the corresponding P from the storage unit. i P of the original data i A real result, and then based on the P i Calculate P based on actual results. i This accuracy rate is derived from a series of inferences. In the data relabeling step, P i The inference result corresponding to at least one of the original data does not match the actual result.

2. The training method for the image recognition deep learning model as described in claim 1, characterized in that, This first data selection step further includes: The T original data points are divided into multiple original data groups, and then a portion of the original data is selected from each of these original data groups to integrate the N data points. i One set of raw data.

3. The training method for the image recognition deep learning model as described in claim 1, characterized in that, M i The following conditions must be met:

4. The training method for the image recognition deep learning model as described in claim 1, characterized in that, P i The following conditions must be met:

Images